How to measure and build trust in AI experiences with eye tracking
Editor's note: This article is an automated speech-to-text transcription, edited lightly for clarity. To view the full session recording click here.
One of the organizations that presented during the 2025 Quirk’s Event – Virtual Global was Tobii. The organization sponsored the event at the bronze level.
Sylvia Knust, director of insight research at Tobii, presented on how eye tracking can be used to build trust in various AI experiences. Knust provided arguments to average objections to eye tracking usage, video examples of how Tobii has used eye tracking to build trust and more.
Session transcript
Joe Rydholm
Hi everybody and welcome to our session, “Building trust in AI experiences with eye tracking.”
I'm Quirk’s Editor, Joe Rydholm. Thanks so much for joining us wherever you are.
Just a quick reminder that you can use the chat tab if you'd like to interact with other attendees during our discussion. And you can use the Q&A tab to submit questions to the presenter, and we'll get to as many as we have time for during the Q&A at the end.
Our session is presented by Tobii. Sylvia, take it away!
Sylvia Knust
Alright, thank you so much, Joe, appreciate it.
Thank you everyone for joining. I'm Sylvia Knust. I'm director of professional services for both the U.S. and EMEA for Tobii. I'm here today to talk about building trust in AI experiences with eye tracking.
So, happy to be here and I see there's a question of where people joining from. I'm actually out of the north woods in Wisconsin right now, so if there is anybody near me, let me know. I'd be curious to see that. Let's get on with things.
It's hard to believe, but I've actually been involved now with eye tracking for nearly 30 years. It's been exciting to watch how it evolved from when it was first used, mostly for military experiences or medical research. Then to packaged research and to web research when the internet got going, way back when it was still static, and there were banner ads that we were testing to see what kind of ads people noticed.
But now we have these hyper personalized experiences, these dynamic digital ecosystems across multiple surfaces, multiple platforms with new features, new designs, developing at breakneck speed, driven by AI, spearheaded by evolving AI design.
Eye tracking has had to really change a lot in response to the new environment, the new ecosystems we work in, the technology, the methodologies, the applications, but it has changed and aggressively so, especially in the last year. And that's what I'd like to talk to you about today.
So, let's first start with what is trust in AI design, right?
Trust is the user’s belief that the system that you were working with is clear, credible and it will support your goals without confusion, hesitation or doubt. Good design keeps users moving forward with confidence.
If you think about that customer journey, AI is creating, enhancing features along the way to help us work faster and do more ideally with more confidence. But it only works if people notice it. It only works if people process it the right way after they notice it and if they can trust it. And it's difficult because trust forms in just seconds and it can break even faster. It can be really hard to rebuild.
So, how do we research and design for this? How do we identify those critical moments? How do we make sure that we've optimized that design to that AI experience to drive trust so that it is perceived as relevant and valuable and people move forward with it.
I want to take a closer look at some of these critical moments. What needs to happen for the customer to move forward with confidence rather than disengage or even worse, lose trust and respect for that brand, product or experience.
First, most importantly, they have to notice it. If they don't see it, it's never going to be experienced. You're never going to have a reaction. But even if they do see it, that does not necessarily mean they'll process or respond to it.
There are stages that our brain moves through very quickly that we're not even aware of.
First, we think about what is it? Is it worth my attention? Is it even worth me pausing longer than just 60 milliseconds to actually pay attention to it and process it to see what it means. Right now, it's in your short-term memory and your brain is making decisions really without you being aware whether or not it's worth becoming aware of and processing it or whether to stay or go.
Essentially, you've got a brief glance, do I stay? Do I go? Do I keep looking at it?
Okay, let's say we've decided to stay. We look at it a bit longer to see if it adds value in the moment.
Do I understand what it is? Does it have meaning or value to me? Do I understand what it's telling me to do and do I trust what it's telling me to do, right?
So, you have to see it. You have to trust what you see. Somehow you decide, ‘yes, this is of value to me,’ and then trust that next step for it to move forward. And then we get that response, either glance at the next element that's important to you, either a click, maybe it's just a thought, a reaction, a change in perception, but this is kind of the process, these micro moments that we have to go through.
AI is all around us, but I would argue that the theme is pretty much the same, right? For any vertical and application. The goals are to get noticed across these channels and through the noise. To add value in that moment and keep in mind the fast-moving world where we are easily overwhelmed. And to drive positive action either towards that brand, product or experience and to build trust along the way with each step.
So, what does eye tracking do? How does it help us? What can it do for us?
It captures moments in real life and allows us to study and understand them in a way that no other research methodology really can. What it captures is real solid observable data points like where we're looking, how we are processing that information and what we decide to do next.
Let's take a look at this particular video.
[Plays example video starting at 5:11]
This is a sample eye tracking video. You can see the environment from the shopper's point of view. You can see in real time where they're looking, how long they look and what they miss. You can fallow along. Notice how they went real broad. First they're just navigating the area. Now they're zooming in on a shelf, their particular area. Now they're comparing different product groups and gradually they narrow in.
Now these are all micro decisions they've made along the way. 'I'm going to move in that direction. I'm way going to scan this shelf. I'm going to narrow in on this area. I'm going to just scan across these particular products.’ So, we see that whole process in that customer journey and all these little micro decisions they make, based on what's noticed, how it's processed and the next step that they take from there.
This is what we're going for. I want to take a couple minutes now just to talk about eye tracking research in general, but specifically how it fits into today's AI-driven landscape.
There are a few misconceptions I want to talk about. I still encounter these today. Now granted, they were all based on legit problems with eye tracking over the years. If you go back 15, 20 years, even 10 years, there were some of these rules or guidelines applied.
There was a time when you had to sit totally still, you weren't allowed to move much like, ‘okay, sit at your desk and we can eye track you as long as you don't move your head.’ That's changed a lot.
There was also a guideline that have to be in a controlled environment and a central test facility. Okay, sorry, but we can only do it for computer screens and nothing else.
We have to use static content or standardized content or prototypes. That has still been, in the last few years, if you want to have standard heat maps where you aggregate all the data onto one background. But even that is not necessary anymore because today testing prototypes just doesn't give you the same value as people experiencing their favorite apps the way they have them set up. We want to have those personal experiences and we can still do heat maps, we can still do metrics.
For a while it said, ‘hey, one surface only is either going to be an app or a computer screen or a digital science.’ We can have all the surfaces you want, all the different channels.
Also, feeling that too much effort, too little value, that it's difficult to add to research. That is definitely not the case anymore.
It has never been so easy to add eye track into your research. Literally all you have to do is put eye tracking glasses on. You can set it up any way you want. Then just put the glasses on and you're eye tracking. You can watch eye tracking live. You can watch those recordings.
Feeling there was only heat maps or video output, also not the case anymore, and that it was too slow to get results. That is still an element, but it's gotten so much faster because just the capabilities are so much more that you can do it very, very quickly. Some outputs are now instant. Others can be processed very quickly based on the goals that we have.
I also want to just touch on the topic right of eye tracking research versus AI in research. I think that's an important topic to think about and AI can be very effective.
Now, one of the challenges is that it tends to use past behaviors, right?
It's everything we already know and it models and it predicts the next behaviors. What's a good probability of happening?
But it's based on essentially what we already know. It's kind of already been established.
Where eye tracking is actual attention, actual behaviors in that moment, tracking how these change every day in response to environments that are changing incredibly quickly. And that's something that I've definitely seen. This goes way back to when they first had ads on the internet and people's attention patterns were changing based on the banner ads, banner blindness certain ad designs, ad blindness. We change in response to our environment.
So, as there's new elements out there, we respond to them differently. So, you need to keep testing these things to how they change.
The other benefit of using eye tracking with this research is it allows you to dig into the “why” of human behavior and define opportunities that we aren't fully aware of. That's really what we're going to dig into quite a bit more as we go along in this presentation.
Really quickly because I want to make sure I have time to dig into our use case today. But just touching on these, the different systems and applications available today.
Advanced screen testing and wearables. You can literally test anything anywhere, except under water, I suppose, but everything else you can do.
Fielding and complex environments. That has definitely been a challenge, but we are now surrounded by so many digital concepts, platforms and services. We can do all of that no matter how dynamic and interactive.
The third bullet, really important, super precise and high speed with the attention data. Because we're learning, and I've learned over the years now, our attention span is getting much faster. Our reaction time is getting much faster. We're making decisions much faster, especially whether to go or stay, reject or accept type of thing. It used to be we could just look at these general things like general dwell time, but now we're looking at saccades, fixations and 60 millisecond periods of time. Are you going to stay or not?
So, it's getting much more precise and our analysis processes have gotten much faster. They have to because we have to keep up.
So, moving along. Here, I'm going to get into a little bit into some outputs. Most of you are likely very familiar with traditional heat maps. For today's use case, we're actually going to look at or to kind of get into this use case, the scatterplot.
This is a visual pattern for one person. It can be lots of people, but right now we're focusing a little bit more qualitative here. A little bit more on each individual to understand trust because its such a nuanced personal experience, even though it does reflect bigger patterns.
This is a visual pattern for one person over 20 seconds, each dot is a fixation, and the eyes have paused. Basically one dot is where the eyes have paused long enough to process the information that's there and make a decision of whether to stay or go. So, not processing enough to say, ‘oh, do I trust this or does this add value?’ This is just, ‘do I care enough right now to stay here or not?’
Consciously, you're not even aware of that to be honest. If you were to be asked what do you remember seeing, you'll just remember seeing things where there's lots of clusters of dots that you actually spend some time on.
That first dot is just, ‘am I going to stare or go?’ And that's what we look forward to see if they are even paying attention.
I noticed there's also some things that are missed entirely. So we do want to look at that fixation, but that fixation alone indicates only that initial processing that is in short-term memory. Only this glance will not have a chance in long-term memory or to impact the user in any meaningful way unless they continue to fixate on that specific element.
We see several dots where we see layers of dots. When we see more of them that is an indication that they're pausing, they're reflecting and they're processing at a deeper level. But that prompts the next question.
When we see those numbers of fixations on that element, is it because they like it and it adds value or is it confusing and difficult to understand? Are they having to reread it? What's going on in their mind at that time? And we want to get at that. We want to know what they're thinking about at that point. We'll get to that in a moment.
But we do see clearly here is the flow of attention, the behavior after each fixation. Do they stay? Do they go? Where do they go? What do they look at next? What do they click? How are they continuing the journey and is it the way we want them to? Is it in the right direction?
If you think back to our original list of goals, getting noticed which elements get noticed, which then where do they stay and where do they feel that it adds value? Does it drive positive action?
And we've seen some interesting patterns here already of like, yep, noticed this, considered it, rejected it, missed some entirely, ignored some entirely or rather deliberately. When you see patterns like that that cut that close around an area, you're like deliberately ignoring that.
Okay, so moving on. With eye tracking, you can do a lot. I could talk for ages on all these different things, but I'm going to focus on some very particular outputs today and specifically for AI research.
But as a review, all of these apply and it kind of goes from yes, you have to do a little bit of analysis and processing, but the instant is the live viewing, right? Just live eye tracking sessions. So basically, now you can see the task, the world from the point of view of the participant, see where they're looking.
Literally, it's so easy to add any research you're doing. Have them put glasses on and you'll see the live output. You'll see the recording like that shopping the video that we just saw, you'll see that.
So, already you can see the world from the participant's point of view, how they're looking at things and start thinking about what are the questions that are coming up in your mind that you want to ask. Like, ‘oh, you paused on this, you reached for, but then you didn't. Can you tell me what you were thinking? What was missing?’
Let those questions start evolving, and those are then questions that you can kind of enhance and refine your exit or your in-depth interview with.
Then we can also use the video during the playback. That's what we want to focus on today is this what we call a Retrospective Think Aloud.
The idea is that when they're doing their tasks, when they're shopping, using an app or shopping on a website, you'd love to know what they're thinking.
What are you thinking? What are you feeling? What's your experience here? Internally? We'd love to know that, but if you ask them to talk it out loud, it changes everything. It changes their whole experience with it. Now they're performing for you. They're not feeling and just doing it naturally.
What we do instead is we play back the video afterwards. We let them watch that video. We take a moment and say, ‘okay, this is you, this is you shopping that dot is your eye movements. What are you thinking about this?’
They are surprised at how fast their eyes move. They're surprised at all the things they looked at and don't even remember looking at. They're surprised when you dig a little bit and ask questions of like, ‘you glanced at this, but you reached for something else. Why?’ They're like, ‘oh yeah, because I wasn't sure if this was the right thing,’ or the other way around.
So many times, they'll take a package, they'll reach for it as they're looking at a different package. And what that means is they're not sure. They think this is the right one, they want to believe it's the right one, but they're also seeing this other package and they're thinking, ‘maybe I should take that one.’ When you see that disconnect between what they have their mouse hovering on and what they're looking at, there's something off here and it's an opportunity to dig in and understand.
Other visualizations, heat maps, we can do those. Opacity maps, gaze overlay videos to highlight behaviors. The scan path and scatter plots, like what we just talked about. Of course, all these metrics, quantitative metrics, a variety, I mean there's just so many we could go into, but I don't have time today. But feel free to follow up with questions on those for sure.
Okay, so we're looking at specifically the video replay and how to leverage that. We're going to go over this in the stages.
First, this is what I would encourage you to do when you're watching live videos, and you have the interview coming up afterwards.
Think about the scatterplot. Remember how we saw an isolated dot that said, ‘okay, they're paused, they notice it, they're deciding whether or not they should stay or go.’ Then there's a bit of a cluster, ‘okay, now they've decided to stay and they're spending more time on it.’
What is the feedback? How are they thinking? What are they thinking about it? Are they confused or is it clear?
We can even look at that scatterplot, and if it's a messy scatterplot, that suggests confusion or is it like, ‘hey, they're reading and they're understanding and processing.’ So, we can go to that level.
So, think about that scatterplot. Now look for these patterns in this video as we watch it.
Okay, so I'm going to play it and here's what I want you to do. I want you to imagine that you're sitting in the back room of a viewing room with clients and stakeholders. You're watching a live view of this online shopping experience, and you want to understand how they are reacting to the AI features.
[Plays example video starting at 16:40]
So, the red dots show us where they're moving, you see their scan, they pause, okay, they just pause again on the AI. Now they're reading it and you're like;
‘I wonder what they're thinking about right now when they read it. Are they going to click on more?
Oh, well, they didn't. Interesting. Why not? Why didn't they click into the AI recommendations that can be listed?’
Now they're looking at this particular look. See the mouse is now on that item, but they're looking at something else. So that's interesting to me. I'd be like, ‘wow, how come you think you're about to click on that? You look at something else and now you go to the filter. So, you didn't click on it. Why not? You didn't think it was the right thing. Why not? Can you tell me a little bit more about that?
So now they've gone to the filters, but they didn't select on a filter. Why not? What was missing?
Now they're typing in something a little bit different.
You can kind of see as I sit by there and with your stakeholders, you're sitting back there watching this and you can talk about these things saying, ‘okay, now they're spending more time looking at the AI, not like they kind of want something there and maybe they're going to now they did click into it. Okay, cool. But they didn't click more. They just clicked a little bit more. Now they're going to go all the way. Okay, they did.’
But the point is you with your stakeholders in the back with clients, who likely will know more and have more expertise about whatever it is they're trying to design, and you have the research expertise and the attention research expertise. You can have those conversations and figure out what those questions are that you want to prioritize in that interview based on this video.
So that's kind of the first step. Think about that scatterplot, the different patterns, the short fixations, the longer fixations, the movements, seeing those in the video, thinking about what questions you want to ask and let that guide your interview. Let that guide, direct and refine your interview process.
So, that's kind of building us up to our case study. We'll want to go a little bit deeper, essentially using that Retrospective Think Aloud.
We've kind of built up our tools here and now we're going to try it. Here, the research objective was understanding how users develop trust in AI generated recommendations. So literally that video we just looked at. What is the reaction to those AI experiences when it brings up recommendations of particular products and brands? How can we continue to improve those designs?
So, it's challenges, right?
Hyper personalized. We want people to use their own accounts, their own devices, not prototypes because it's already refined to their experience. Those critical moments are quick, they're brief, it's difficult to capture and unpack. We're dealing with trust in these shortened moments, and there's a tendency of bias towards what the moment participants express in that qualitative feedback. They often want to share positive things, so it can be difficult to get them to really give you something more concrete.
So, how do we unpack those things that are difficult to observe and define?
Really quick, again, the objective for design was optimizing those moments of truth with AI generated recommendations within those e-commerce experiences. The desired output for the client was to really optimize those moments in time, the format for AI integration, promote the noticeability and engagement. And especially what are those cues that drive trust and inhibit any doubt?
These days with digital content, there's more mistrust and there's more doubt uncertainty around that. So, we intercepted people in high traffic areas. We use glasses. Literally, we just stopped people in high traffic areas, and we're like, ‘hey, you want to participate in this research? Bring up an app that you use for shopping, put these glasses on.’ Then we did Retrospective Think Aloud.
This should look familiar. This is the same video. We're going to use just the same one because it's more familiar. Again, some of the kind of rules or guidelines. You want natural user interactions.
No talking out loud, no explaining what you're thinking, just super natural. Just put these glasses on, do your thing. We observe those attention patterns and all those clicks and behaviors in real time. And as you're observing it, we let those crystallize. Remember those questions we had in our mind as we were watching the video.
Include those stakeholders. It is so helpful to have the stakeholders involved. You can do this any way you want, but have them watch and have them share. It's amazing what you can learn from what's important to them and how they think about their products and designs.
Prioritize those key moments and questions.
Then play this back during the interview. So, pick your points in time and play those back during the interview.
Here's an example of what the outcomes are. These are some recordings and snapshots. You see the scatterplot again, and you're going to see a video here.
[Plays example video starting at 21:12]
Again, this was on mobile devices, very similar to that video that you saw on the PC.
We are watching them reading through the AI and what you'll notice here is these multiple patterns. They go back and reread it again and then they go back and reread it again and again. And we're like, why?
So, then in the interview, we literally played this back, and we said, ‘hey, we noticed this. We noticed that you kept going back and rereading it a few times. Why?’
We wanted to know was it confusing? Was it hard to understand? What's happening here?
And they said ‘no, it was hitting the key points. And I skimmed at first and I was like, wow, wait a minute. It touched on exactly the points I wanted to and then I would remember other things I wanted to know. And I went back and I was like, oh yeah, they touched on it here.’
So, the first time through, they didn't read everything yet. They were doing an initial skim. Is it worth my time?
They went through it again. They said, ‘yeah, it is worth my time.’ And they went through the third time, they're like, ‘yeah, actually it's got me figured out perfectly. I'm going to find out what these recommendations are.’
So, then they went to the recommendation page over here on the second one. And at the top saw these recommendations. They swipe through a couple.
Then what you notice here is actually the third recommendation. And then they looked to the right. You can see that the dots, the fixation, they're trying to swipe again, they want to swipe for the next recommendation and the next recommendation. They wanted to see more than just the top three. This wasn't enough. This didn't give me enough confidence that this was actually designed for me because they're used to kind of the sponsor recommendations.
We always see two or three sponsor recommendations and it's like, ‘well those are paid. They're not thinking about what I need, what I want.’ So, there's quite a bit you have to work towards to rebuild that trust.
Then on this third example, you can see they briefly considered several other products. Then they spent a ton of time just on this particular one. Why? And that's the one they ended up clicking on.
But they had their mouse on this first one first. Then glance with these other two. Then looked over this one. Again, why?
As they watched their own eye movements, they kind of relive the experience. They can tell you those nuanced decisions they made, that at the time they weren't even aware of. And if you didn't play the video, you might get some broad feedback. But this gets super nuanced, super tight, super focused on specific design elements and experiences.
So, how to move forward with that. I'm super excited about eye tracking. I think it adds so much value to the research that we do. And I'll share with you a few things over the 30 years to highlight just the value I've seen in it and why, today especially, it's so important.
What I've learned over the last year, is that attention spans really are getting shorter and shorter. The eyes move so quickly, that's usually the first thing people will say when they watch an eye tracking video themselves. They also are surprised that we look at, and we process so much that we are not consciously aware of. But if we look at the eye movements and the scatterplots and the videos, we can see and understand a lot more about what our brain is doing that we aren't necessarily articulating without seeing that and understanding that.
The other thing is just about general research. You kind of touched on a lot of these already. Tunnel vision and bias are real for participants and researchers.
As soon as you start designing a study, and I know sometimes you have to, but as soon as you put parameters on that you're already introducing bias where you want to try and keep things a little bit more open, keep it real. I cannot stress that enough.
Natural environments, natural behaviors, whatever they normally do, use their own phone, use their own apps, keep it natural.
Perceptions and behaviors are super quick to change. I can't stress how important it is to test often.
For me, doing research without eye tracking is like flying blind. It's really hard. There's so much going on that I don't understand or don't know because I don't see where they're looking.
You can use it in all these different applications, observe the world through the eyes or just that alone can really do so much to understand how people are engaging with the world around.
You can break a lot of assumptions. It can inspire your designers.
You can use it to probe and get deeper at those really tricky questions. Define those key moments that are so important and take a moment to really dig into that and understand how to design for and develop guidelines specifically for AI research.
It provides that first person data, that real time focus and attention, like what's happening right now. Highlighting those critical decision-making moments. And you can use that to unpack that moment and the impact on trust and how to design better to create trust.
So, that is it. Happy to connect more. I will take some questions. So I want a little longer, but we have a few minutes for any questions.